= Expected utility of a Gaussian location-scale family
Let $X=\mu+\sigma Z$, where $Z$ has the <standard normal distribution> and $\sigma\geq0$. Suppose $U:\mathbb R\to\mathbb R$ is increasing, concave and twice continuously differentiable, with sufficient integrability to differentiate its Gaussian expectations twice and integrate by parts. A convenient sufficient bound is $|U(x)|+|U'(x)|+|U''(x)|\leq C e^{c|x|}$ for fixed constants $C,c$. For $f(\mu,\sigma)=\mathbb E U(X)$,
$$
f_\mu=\mathbb E U'(X)\geq0,\qquad f_\sigma=\mathbb E[ZU'(X)]=\sigma\mathbb E U''(X)\leq0.
$$
The middle equality uses <Gaussian integration by parts>. If $U'>0$, the mean derivative is strictly positive. At $\sigma=0$, the scale derivative is zero. Allowing negative $\sigma$ leaves the distribution unchanged but reverses the sign of that derivative, so the monotonicity assertion is stated on nonnegative scale.
Furthermore,
$$
\operatorname{Hess}f=\mathbb E\left[U''(X)\begin{pmatrix}1&Z\\Z&Z^2\end{pmatrix}\right]
$$
is negative semidefinite: its quadratic form on $(a,b)$ is $\mathbb E[U''(X)(a+bZ)^2]\leq0$. Thus $f$ is jointly <concave> in mean and scale. Concavity also follows directly by applying concavity of $U$ pointwise to affine combinations of $\mu+\sigma Z$, without assuming the differentiability formulas when their integrability fails.
Back to article page